Artificial Intelligence for Predicting Active Lesion in Multiple Sclerosis from Non-contrast MRI
Artificial Intelligence for Predicting Active Lesion in Multiple Sclerosis from Non-contrast MRI
AmirAbbas Amini,1,*Raheleh Kafieh,2Azin Shayganfar,3Zahra Amini,4Leila Ostovar,5Somayeh Haji Ahmadi,6
1. School of Advanced Technologies in Medicine, Medical Image and Signal Processing Research Center, Isfahan University of Medical Sciences, Isfahan, Iran 2. Department of Engineering, Durham University, Durham, UK 3. Department of Radiology, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran 4. School of Advanced Technologies in Medicine, Medical Image and Signal Processing Research Center, Isfahan University of Medical Sciences, Isfahan, Iran 5. Department of Radiology, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran 6. Department of Radiology, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran
Introduction: Multiple sclerosis (MS) is a disease that affects the central nervous system. In this disease, the myelin covering the nerve fibers is attacked by the body's immune cells and creates lesions in the brain, these lesions are classified into two types, active and inactive. All these lesions can be identified in fluid-attenuated inversion recovery (FLAIR) type MRI images, but it is impossible to distinguish whether they are active or inactive. Therefore, in order to detect active lesions and control the disease, MRI imaging with gadolinium-based contrast is used, but since the long-term deposition of gadolinium in various tissues can cause complications for the patient, it is important to investigate alternative methods. The purpose of this study is to investigate the deep learning method as one of the methods based on artificial intelligence in the diagnosis of active lesions without the use of contrast agents.
Methods: Our data were collected from 130 patients with Relapsing-Remitting Multiple Sclerosis in four different imaging centers in Isfahan City. Firstly, the lesions were identified by radiologists and classified into active and inactive categories. These lesions included a total of 9097. Then, each lesion was separated as an ROI from the FLAIR sequence MRI scans to be used as the input of the artificial intelligence network. Next, three deep learning networks including a convolutional neural network (CNN) as the main designed network, and two transfer learning networks including VGG19 and Efficient NetB0 were designed and trained to distinguish active from inactive lesions. Finally, the statistical results obtained from each network were calculated and compared with each other.
Results: For our designed CNN, the average results of precision, recall, and F1 score in 5-fold cross-validation for active and inactive classes were 0.77, 0.99, 0.87 and 0.98, 0.70, 0.82, respectively. These values were obtained for the VGG19 network 0.68, 0.87, 0.76, and 0.82, 0.60, 0.68 respectively, and for the Efficient NetB0 network 0.67, 0.94, 0.78 and 0.90, 0.57, 0.67 respectively. Also, the values of accuracy and areas under the receiver operating characteristic curve (AUC) were evaluated for these 3 networks, which were 0.85 and 0.94 for our designed CNN network, 0.73 and 0.79 for the VGG19 network, and 0.74 and 0.81 for the Efficient NetB0 network, respectively.
Conclusion: Our results show that the use of deep learning networks as one of the methods of artificial intelligence can accurately detect active lesions of MS disease. This method can avoid the side effects of contrast injection.